Random Forest Training:

// define labels to be trained std::vector labels;

// for NYU depth dataset v2 this could be… labels.push_back(21); //

wall | labels.push_back(11); // floor labels.push_back(3); // cabinet labels.push_back(5); // chair labels.push_back(19); // table labels.push_back(4); // ceiling

// load training data from files ClassificationData trainingData; trainingData.LoadFromDirectory(“myTrainingDataDirectory”, labels);

// define Random Forest // parameters: // int nTrees // int maxDepth (-1… find depth automatically) // float baggingRatio // int nFeaturesToTryAtEachNode // float minInformationGain // int nMinNumberOfPointsToSplit Forest rf(10, 20 , 0.1, 200, 0.02, 5);

// train forest // parameters: // ClassificationData data // bool refineWithAllDataAfterwards // int verbosityLevel (0 - quiet, 3 - most detail) rf.TrainLarge(trainingData, false, 3);

// save after training rf.SaveToFile(“myforest”);

// Hard classification, returning label ID: std::vector featureVector; // assuming featureVector contains values… int ID = rf.ClassifyPoint(featureVector);

// Soft classification, returning probabilities for each label std::vector labelProbabilities = rf.SoftClassify(featureVector);

---

Directory structure of training data:

mytrainingdirectory - 0001.data - 0002.data - 0003.data … - 1242.data - categories.txt

-> for every label (the example above has 1242 labels) there is a corresponding data file containing all feature vectors of this label. such a file looks like this:

2.46917e-05 0.000273396 0.000204452 0.823049 0.170685 0.988113 0.993125 3.20674e-05 0.000280648 0.000229576 0.829844 0.207543 0.987969 0.992765 3.73145e-05 0.000279801 0.000257583 0.831597 0.235013 0.987825 0.9925 ……….. ……….. ……….. ……….. ……….. ……….. ………..

Each row: One feature vector (here 7 dimensional)

IMPORTANT: Values separated by spaces, vectors separated by newline. COLUMNS MUST ALL HAVE THE SAME CONSTANT WIDTH!!! CODE IS ASSUMING THAT FOR SPEED UP! (here width is 12)

The file categories.txt is necessary to define the label names and has a simple format like this: 1 floor 2 wall 3 ceiling 4 table 5 chair 6 furniture 7 object

Original page: https://github.com/strands-project/v4r/blob/master/modules/ml/README